You are on page 1of 10

Wear 426–427 (2019) 1702–1711

Contents lists available at ScienceDirect

Wear
journal homepage: www.elsevier.com/locate/wear

Automatic classification of severe and mild wear in worn surface images T


using histograms of oriented gradients as descriptor

C. Gonzalez-Ariasa, C.C. Viafaraa, , J.J. Coronadob, F. Martinezc
a
Corrosion Research Group, Metallurgical Engineering and Materials Science School, Universidad Industrial de Santander, Cra. 27-Cl. 9, Bucaramanga, Colombia
b
Research Group of Fatigue and Surfaces, Mechanical Engineering School, Universidad del Valle, Cl. 13 #100-00, Cali, Colombia
c
Biomedical Imaging, Vision and Learning Laboratory (BIVL2ab), Systems and Computing Engineering School, Universidad Industrial de Santander, Cra. 27-Cl. 9,
Bucaramanga, Colombia

A R T I C LE I N FO A B S T R A C T

Keywords: The main objective of this work was to develop a novel computational strategy that is able to predict wear
Wear monitoring regime operation in worn surfaces. The image data were taken from worn surfaces images of cast iron specimens
Image processing techniques subjected to abrasion wear tests. These images were classified into two groups, identified with the severe and
HOG descriptor mild labels, according to the wear rate results found during the wear tests. The surface features of worn surfaces
Abrasive wear regimes
images were coded as a dense Histogram of Oriented Gradient (HOG) descriptor and thus classifier models were
Mild wear
Severe wear
herein implemented to obtain a learning model of wear severity. Gaussian Naive Bayes, Decision Tree and
Random Forest were the classifier models used, which span the family of classifiers from fast to robust im-
plementations. An evaluation of the classifier capacity to identify those images corresponding to the severe and
mild wear regimes was made by following a k-fold cross validation strategy. The qualitative characterization of
worn surfaces images through the HOG computation and the application of classifier models allow predicting
well whether a mild or a severe abrasive wear regimes operated. The proposed approach achieves more than
80% of accuracy in almost all HOG configuration and for the different classifiers herein evaluated.

1. Introduction inspectors [4–8]. The condition monitoring is usually made by ac-


quiring images from worn surfaces or from wear debris surfaces. Then,
Wear problems in industry applications involve not only mass losses these images are analyzed for extracting the surface characteristics
of mating surfaces but also time, energy and material losses, which using a broad range of image processing techniques. Some of the most
result in high economical costs of the corresponding maintenance and common image processing techniques which have been generally ap-
reparation activities [1]. For this reason, there is a demand for an im- plied for monitoring wear involve thresholding [5–12], edge detection
proved reliability of systems in order to reduce shut-down and un- or operators estimation [5,6,9,8,11,14–16], fractal analysis [4,16–18]
availability of machine components [2]. An optimized reliability of and stitching [12,19–21].
machine elements could help in preventing failures by means of esti- Thresholding consists on group similar pixels by establishing dif-
mation of the useful life and the probability of failure. From this point ferent threshold levels, and associating such groups to different classes
of view, wear inspection is a fundamental issue to be developed and [22]. This is the main processing technique used for image segmenta-
implemented in engineering applications, also incorporating smart tion, for instance, for partitioning worn and unworn regions of a sur-
systems technology [3]. face, as well as for determining the geometrical profile of a surface. This
The wear condition monitoring through the image processing explains why thresholding has been commonly used during the condi-
techniques has been extensively investigated considering its many ad- tion monitoring of cutting tools [5,7–10], where the flank and crater
vantages [4–21]. This vision-based monitoring can be automated and wear must be differentiated from the unworn surface of tool. Moreover,
allows obtaining an online, non-contact, accurate and cost-effective thresholding is used as an image pre-processing technique for reducing
inspection of worn surfaces [4]. Additionally, it is widely accepted that noise [6]. This strategy is however sensible to color changes, perspec-
imaging systems are promising to support and even replace human tive of capture and illumination variations.


Corresponding author.
E-mail addresses: camilo.g.ct@gmail.com (C. Gonzalez-Arias), cviafara@saber.uis.edu.co (C.C. Viafara), John.coronado@correounivalle.edu.co (J.J. Coronado),
famarcar@saber.uis.edu.co (F. Martinez).

https://doi.org/10.1016/j.wear.2018.11.028
Received 4 September 2018; Received in revised form 15 November 2018; Accepted 27 November 2018
0043-1648/ © 2018 Elsevier B.V. All rights reserved.
C. Gonzalez-Arias et al. Wear 426–427 (2019) 1702–1711

The edge detection technique is able to recognize edges in an image Table 1


by estimating operators such as gradients and directions [22] or tex- Description and hardness of the sample materials [25].
ture-based operators [8]. The selection of the operator defines the Symbol Sample materials Bulk hardness
computational cost of the image-based technique. More robust opera-
tors could still be employed for image segmentation [8]. In a worn [HV98N ] [MPa]
surface the quantity of detected edges can be correlated with the wear *
Wa WCI with austenitic matrix 694 ± 16 3670 ± 85
intensity [6,8,11], while in wear debris this technique can identify Wm WCI with martensitic matrix 777 ± 14 4109 ± 74
many geometrical parameters [14]. A main limitation of edge-based W14 WCI with 13.8 wt% Cr 966 ± 30 5108 ± 159
characterization is that much of the times also recovers noise gradients W24 WCI with 24.4 wt% Cr 1062 ± 27 5616 ± 143
that could represent statistical significance, leading to errors in re- M MCI* 493 ± 8 2607 ± 42

presentation.
* WCI: White Cast Iron, MCI: Mottled Cast Iron.
Fractal image analysis has been also used in different approaches for
monitoring wear surfaces, since it carries out a numerical character-
The aim of this work was to develop a computational tool for image
ization of the surface morphology [16–18]. Fractal characterization
analysis and understanding of worn surfaces exhibiting severe and mild
usually assigns numerical values to heterogeneous surfaces which have
abrasive wear. For this, Scanning Electron Microscopy (SEM) images of
been controlled by chaotic processes such as wear [17]. As a result,
worn surfaces obtained by laboratory abrasion tests published else-
some works using this technique have been found correlations between
where [25] were analyzed using a Histogram of Oriented Gradients
fractal values and the erosion wear mechanisms [16–18]. Also, the
(HOG) descriptor, together with different image classifier models, ty-
fractal analysis overcomes problems of local noisy by computing dif-
pically used in machine learning applications. An evaluation of the
ferent scales of images, preserving main information along the scales.
classification rate for each model was carried out by applying a cross
For in situ detecting the evolution of a wear process, the stitching is
validation strategy.
the most commonly used image-based technique. This technique com-
bines and matches images of a surface obtained in different periods of
time and then detects the topographical changes occurred [21,12]. It 2. Experimental procedure
has been employed in the characterization of wear mechanisms in
sliding wear [19,20] and in rolling-sliding contact tests [21,12]. 2.1. Selection and classification of images
As mentioned above, wear monitoring using image-based techni-
ques have been applied with different kind of analysis, such as cutting The selected images for analysis in this work were obtained from an
tool wear detection [5–10], wear particles [13,14,16] and laboratory investigation about the effect of the hardness of abrasive particles on
wear tests characterization [4,11,12,15,17–21]. Some works found that the mild-severe wear regime transition of cast irons [25]. Abrasive wear
wear progress of analyzed surfaces can be identified by using the image tests were carried out varying both the hardness of the sample materials
processing tools [10,7,14–16]. In particular, Pedersen reported that a (H) and of the abrasive particles (HA ). Detailed information of the wear
three-stage wear pattern was possible to be recognized during the tests and the tested materials can be found elsewhere [25]. In Tables 1
cutting tool wear monitoring [7]. Moreover, those studies analyzing and 2 are showed the description and hardness of the sample materials
worn surfaces images reported, for the most part and including the and of the abrasive particles, respectively.
studies of tool wear monitoring, a quantitative correlation between the During the abrasion tests the sample materials were rubbed against
wear behavior and the numerical parameters obtained from the pro- each one of the abrasive particles, from which several hardness ratios
cessed images [6,10,7,8,17,19,4,11,18,20,15]. On the other hand, a few (HA/ H ) were obtained. The wear rate as a function of the hardness ratio
works only obtained a qualitative correlation between the image ana- can be observed in Fig. 1, as usually made in the literature [1]. In this
lysis results and the features of worn surfaces [4,21]. In spite of the few graph two main regions were distinguished corresponding to the mild
works executing a qualitative characterization of worn surfaces, this and severe wear regimes. In the mild wear regime, low values of wear
kind of analysis could still contribute for monitoring the wear regimes rate were produced when the HA/ H ratio was below 1.9 [25]. It must be
operation. noted that in the mild wear region, a higher wear rate was achieved
An operation of severe and mild wear regimes has been observed with increasing the HA/ H ratio. However, those highest values of wear
and characterized during the sliding [23] and the abrasive wear of iron rate in the mild wear region were still considered as representing a low
alloys [24,25]. Each wear regime produces well-distinguished features wear rate. At values of the HA/ H ratio higher than 1.9, a region of
in worn surfaces, wear debris and obviously in wear rates. The most severe wear regime was obtained. In this region, an approximately
important fact is that the differences in wear rates between both wear constant wear rate was observed for all conditions.
regimes are from one to three orders of magnitude [23,25]. Ad- In Fig. 2 the same wear rate data are represented in terms of the
ditionally, a transition between severe and mild regimes can be pro- power dissipated by friction. It can be noted that the same two regions
moted by small changes in the operational conditions of the tribological of mild and severe wear are distinguished at the low and high values of
system [23–25]. In the case of sliding wear of steels, a slight variation of friction power, respectively. Similar results have been obtained in the
normal load, sliding velocity or hardness of both sliding bodies could mapping of rail wear regimes and transitions [27]. In the present case,
promote the wear regime transition [23]. For the abrasive wear, a the changes in contact conditions through the increase of the hardness
transition from mild to severe is caused when the material hardness-to- of the abrasive particle lead to a higher dissipation of energy for time
abrasive hardness ratio (HA/ H ) overcomes a specific value [24,25].
Considering that some changes in operational conditions can be found Table 2
in industrial systems [26] and from the viewpoint of machine reliability Description and hardness of the abrasive particles [25].
[2], it is of crucial importance to inhibit a wear regime transition from
Abrasive particles Hardness
mild to severe. In other words, and in terms of industrial applications,
an operation of a severe wear regime must be strongly prevented. [HV0.98N ] [MPa]
To the best of our knowledge, an image analysis technique still has
not been applied for recognition of severe and mild wear regimes. Then, Glass 552 ± 26 2920 ± 138
Garnet 1453 ± 85 7687 ± 450
a condition monitoring by using image processing methods could pro- 2563 ± 124 13559 ± 656
Al2 O3
vide an appropriate tool to detect and diagnostic the wear regimes SiC 3487 ± 322 18447 ± 1703
operation only by an automated recognition of worn surfaces images.

1703
C. Gonzalez-Arias et al. Wear 426–427 (2019) 1702–1711

describe different texture patterns. Then, a computational description


and an automatic classification of wear regimes is very challenging
from traditional global representation of gradients, color-based de-
scriptors or strategies that try to recover specific shapes from surfaces.
For example, the computation of global approaches that quantify gra-
dients in whole images lost local description that could be fundamental
to understand the severity degree. Also, the approaches that only use
the norm of gradient as surface signature are prone to errors because
the diverse textures associated to the same degree of wear severity. In
this work, a local gradient quantification is carried out by computing a
set of oriented histograms along the images, allowing a local descrip-
tion of surfaces. Also, the HOG descriptor achieves a dense and re-
dundant estimation of gradient histograms, which are computed fol-
lowing a sliding window. To remove point gradient artifacts, the
histograms are normalized with respect to a neighborhood. The histo-
gram of gradients gives an idea of main orientations reported into a
specific region, while the dense computation along the image emerge to
Fig. 1. Wear rate of sample materials as a function of the hardness ratio (HA/ H ) be robust with respect to the variability of representation of surface
[25].
features of wear regimes. Hence, such sophisticated gradient descrip-
tion allows to robustly train learning strategies models that achieve a
proper discrimination between mild and severe wear classes.
Specifically, the dense HOG computation starts by dividing an
image into a set of patches with horizontal and vertical lengths of size
c = ch × c v , respectively. Then a local histogram of gradients Hi is
computed to represent each of these ci patches. For a specific region c ,
with gradients ▽c , and respective orientations: ϕ (▽c) , the HOG de-
scriptor computes gradient orientation bins ω as the relative occurrence
of gradient vectors with a given orientation, weighted by their vector
norm:

Hi (ω) = ∑ ▽c
{ϕ (▽c) = ω} (1)

where ω ∈ {ω0…ωN − 1} and ωN is the number of orientations. Each of the


computed bins ω are normalized w.r.t to the sum of the total of norm
gradients on the specific region c . In this case, a smooth surface with
mild wear could be represented as an uniform histogram with equal
Fig. 2. Wear rate of sample materials as a function of the power dissipated by probability for all gradient orientations. In contrast, local c i patches
friction.
that capture surface features of severe wear are represented by asym-
metric histograms with gradient bins that stand out from the others and
unit and then to the action of a severe wear regime. This constitute depict the non-uniformity of the surface. Nevertheless, the observed
other approach used in literature to classify types of wear regimes. Both wear regimes report high variability on gradient patterns, being a very
approaches give a basis to classify the SEM images of worn surfaces. challenging problem (see Fig. 3). For instance, in Fig. 3 (b) there are
A total of 50 SEM images of worn surfaces of six of the hardness some uniform global gradient patterns but with a main direction on
ratios were selected, classified and labeled in according to the two wear local histograms. Finally, a regional analysis of a set of neighborhood Ω
regimes. 29 images corresponded to the test conditions exhibiting se- patches could be achieved by concatenating the respective neighbor-
vere wear and 21 images were associated to the test conditions in which hood histograms bn = {H1, H2, …Hz ; Hi ∈ Ω} and performing a block
the mild wear occurred. The images consisted in a 256 × 256 pixels normalization with respect to the total number of bins of whole histo-
array, with a range from 0 to 255 in grey level and giving 3D data of the bn
grams in a specific region, as shown in: bn = .
bn +ε
surface texture. In Fig. 3 the SEM images corresponding to some of the The whole HOG configuration is illustrated in Fig. 4. From now on
test conditions are showed, where the main differences between surface the notation of p-cq-br will be taken to express a HOG configuration of p
features produced in both wear regimes can be recognized. Mild wear orientations, q patch sizes and r grouped patches in the normalized
promoted worn surfaces with a smooth appearance and low amount of block. The number of orientations, the patch size and the number of
shallow scratches, meanwhile worn surfaces with severe wear had a grouped patches to be normalized constitute the main parameters of
rougher surface with plastic deformation traces, microchips formation HOG descriptor to capture wear patterns in SEM surface images.
and deeper scratches [25]. Wear patterns observed in SEM images have A proper trade-off of such parameters can describe main geome-
already been characterized with a vision-based tool [17]. trical wear shapes within a reasonable computational time to obtain a
prediction about a particular surface. For instance, a small size patch
2.2. Geometrical image descriptor: the histogram of oriented gradient requires few bins in the respective histograms, allowing to describe
(HOG) approach atomic patterns but being sensible to illumination changes and common
image variations. In such case, it is recommendable to make stable the
Wear severity of surfaces can be associated with non-uniform gra- gradient description by grouping a lot of patches and using a same
dient patterns in captured images (see SEM image examples in Fig. 2). normalization. On the other hand, large size patches can include dif-
Nevertheless, SEM images with associated mild and severe wear fea- ferent gradient wear patterns and histograms with more bins are re-
tures report a huge variety of gradient patterns along the surface. For quired. In this work, the best parameters were found experimentally but
instance, in Fig. 2, the three examples of severe wear conditions in general the proposed approach is relatively stable with different

1704
C. Gonzalez-Arias et al. Wear 426–427 (2019) 1702–1711

Fig. 3. SEM images of the worn surfaces for the W14-glass (a), Wa-glass (b), W24-garnet (c), M-SiC (d), M-garnet (e) and Wm-garnet (f) conditions. Note that images
(b) and (f) are quite similar in terms of the gradient patterns, which difficult automatic representation and classification tasks. Also, these group of images evidence
the high variability of wear regimes features.

parameter configurations. This HOG descriptor has been successfully describe regions with HOG because its simplicity allows to understand
used in many context and applications of computer vision, such as the computed patterns from recovered shapes.
object detection, person recognition, among much others. Essentially,
this descriptor is able to robustly describe any geometrical configura- 2.3. Wear image classification
tion by using occurrence statistics of the gradients in local regions. The
local description of gradient histograms and the block regional nor- An automatic classification step was herein implemented to auto-
malization achieve a dense global description of patterns that are pre- matically label each image according to the wear severity. This classi-
sent in a particular image. fication task comprises the automatic labeling of two different classes of
In our particular case, the wear severity can be successfully de- wear regime, severe and mild, expressed as y = {0, 1} , respectively. For
scribed by using the HOG descriptor, since the “abnormal” severe so doing, each image is represented with a HOG descriptor vector as
patterns are stand out in non-uniform histograms captured along the x = {x1, x2 , …x n} , where each x i is a resultant bin feature of the de-
sequence. Other similar approaches has been proposed in this direction scriptor. Then a set of m training image descriptors x1, x2, …xm are used
like the edge orientation histograms, scale-invariant feature transform to build a training model that allows to establish boundaries between
descriptors, and shape contexts, but differ in how geometrical config- the two categories with some marginal error. The classification task for
urations are captured and the number of measures of histograms ob- a particular sample can be modelled as the probability of belong to a
tained in an image [28]. In this exploratory work we decide to locally class yi given a particular HOG descriptor. In such case, the new sample

1705
C. Gonzalez-Arias et al. Wear 426–427 (2019) 1702–1711

n
y^ = arg max ∝ P (yk ) ∏ P (xi |yk = 0)
k = {0,1} (4)
i=1

With y^ as the estimated severity label for a new image given the
HOG descriptor. This classifier constitutes a fast and efficient alter-
native to code independently each of the descriptor features. This al-
ternative can easily be extended to real time applications but could be
sensible to image surface variations.

2.3.2. Decision tree classifier (DT)


An additional alternative herein considered was the classification of
wear images but implementing a non parametric computational tree
algorithm [30]. In this case, the conjunction of HOG features
{x1, x2 , …x n} represents the branches of a particular learning tree and the
labels y = {0, 1} are represented in leaves.
Fig. 4. An illustrative example of the HOG descriptor computation over a SEM In this case, it is assumed that each of the HOG features has a dis-
image is shown. Each region of the grid is the computed patch with size ch × c v . crete domain and the learning tree is obtained by recursively splitting
At each of these patches is computed an oriented histogram. As illustrated in the the defined domain in a way in which groups are formed with the
image, in mild wear regions the corresponding histogram is uniform with low closest features. Then, from a set of training HOG descriptors {xj} j1 … m
magnitude in each of the bins. In contrast, for severe wear regions non-uniform
with the corresponding severity label annotations y = {0, 1} is defined a
histograms are computed. A set of patches bn (red square) are joined together to
set of candidate splits as θ = (j, τm) regarding the HOG feature j and a
carried out a global normalization. Finally, the HOG descriptor corresponds to a
set of whole oriented bins computed along the image (For interpretation of the
particular threshold partition τm . According to the defined maximum
references to color in this figure legend, the reader is referred to the web ver- depth of the tree, a set of thresholds τm in each dimension of HOG de-
sion of this article.). scriptor are learned to maximize the class discrimination. In this work
was defined a maximum tree depth of two for each dimension, taking
into account the larger dimension of HOG descriptor. The learning al-
to be classified is projected to a trained space and the closest label is
gorithm then try to split the space of features in subscapes that contains
assigned to this sample.
patterns of one specific class. At each node Q of the tree is therefore
In the literature there are several algorithms to perform such clas-
defined a partition Qleft (θ) and Qright (θ) . From such division, a “Gini
sification that consider different assumptions to split the space of fea-
impurity” term IG , is introduced that measure how often a randomly
tures and establish boundaries among trained classes [29–32]. The ef-
chosen element on HOG descriptor would be incorrectly labeled at a
fectiveness of such methodologies depend on size, quality and nature of
specific node G (Q, θ) . During training, the Gini term IG is computed by
the data, which difficult the selection of a proper approach. These al-
summing up the probability of each HOG descriptor pi (with label i)
gorithms also differ in the computational time which is correlated with
times the probability of incorrect labels k, being zero when all miss-
the linearity assumption of data and features used for training. The
clasifications point out to a same label. The expression for Gini term is:
wear classification represented from gradient patterns may be highly
IG = ∑k ≠ i pk (1 − pi ) . Hence, the parameters are selected following a
variable, and much of the regions can be shared in both severity classes.
minimization rule, as defined in Eq. (5):
In this work we consider three different classifiers that span the family
of classifiers from fast implementations using weak linear assumptions θ* = arg min G (Q, θ)
θ (5)
to more robust classifiers that try to model non-linearities in boundaries
of classification to perform a more sophisticated classification but with Typically, a global optimization process is carried out to set the
an important increase in computational time. The set of models used in corresponding decision tree from training samples. In this work, a
the comparison are described as follows: hierarchical optimal discriminant analysis, CART strategy, is carried
out by building binary trees until achieve a stable tree description. This
2.3.1. Gaussian naive bayes classifier (Gauss) work implements the algorithm using sklearn library, that consider an
This classifier is one of the most simple and fast algorithm to per- optimized version of the CART strategy. A main advantage of this
form the classification task under strong (naive) assumption of condi- classifier is the description of results and the possibility to visually
tional independence among features of the descriptor {x1, x2 , …x n} [29]. understand the label decision through the feature branches. However, a
Then, the conditional distribution of a particular label given the HOG simple tree can be insufficient to learn the whole variability of data and
descriptor can be modelled as shown in Eq. (2): the prediction can vary according to the number of samples taken to
build the tree.
n
P (yk = 0|x1, x2 , …, x n ) ∝ P (yk = 0) ∏ P (xi |yk = 0)
i=1 (2) 2.3.3. Random decision forest classifier (RaF)
In this work was also analyzed the performance of automatic wear
where each x i corresponds to a particular bin in the HOG descriptor and
prediction by using a set of decision tree classifiers grouped in an en-
yi is the (mild/severe) label. Also in training model, we assume that
semble learning strategy named the random decision forest [31,32].
each of the x i features is associated with the particular class yi according
This classifier overcomes some sensibility problems of decision trees by
to a Gaussian distribution, as expressed in Eq. (3):
computing several trees to take a robust decision.
(
P (x i |yi = 0) =  μ xi , σxii ) In general, DT classifier could have low bias and very high variance
(3)
in results, being strongly dependent of initialization parameters and
Then, in the training step are build the set normal distributions features used in the learning process. In random forest strategy, a set of
 (μ xi , σxii ) for each HOG feature according to each known label. For a independent DT algorithms are trained over different parts of HOG
new image represented as a particular HOG descriptor x , each of the features space to reduce the variability in the prediction. For so doing, a
features is projected to the normal trained distribution and a maximum Bootstrap aggregating strategy is implemented, which consist on select
a posteriori (MAP) rule is implemented to return an automatic label, as randomly a set of training HOG features {x1, x2 , …x n} to built a particular
seen in Eq. (4): decision tree fb . This process is performed several times to obtain B

1706
C. Gonzalez-Arias et al. Wear 426–427 (2019) 1702–1711

different trees with the capability to perform prediction of severity label configurations. Initially, it could be noted that the HOG results are more
yi from a set of reduced HOG features. In order to preserve computa- uniform in the mild wear image than in the severe wear image. This is a
tional efficiency for whole experiments, the proposed strategy was ran consequence of the few features which are found in the worn surface
with a total of 10 trees in the forest (B = 10 ). The final prediction may with the mild wear. On the contrary, the severe wear image exhibits
be carried out by averaging the predictions of individual trees, or taking more and notable surface features such as edges, which result in a
the majority vote as expressed in Eq. (6): particular pattern obtained in the respective HOG computation.
B As a consequence of the variation in the HOG parameters of or-
1
y^ = ∑ fb ∧ y^ = arg max {f1 , …fB } ientation and patch size, it can be observed some differences between
B b=1 (6) the HOG results for both HOG configurations. In the first place, we can
note the higher quantity of small patches in the 4 − c16 − b1 config-
An estimation of uncertainty of the prediction from the family of
uration, which allows to identify with more detail the surface features
computed trees can be also expressed in Eq. (7):
in the image. In other words, and depending on the scale of the wear
B
∑b = 1 (y^b − μ (y^ )) characteristics, this HOG parameter must be optimized. On the other
σ= hand, the quantity of HOG orientations does not seem to have a marked
B−1 (7)
effect. This could be a result of the fact of the worn surfaces have a
This strategy in much of the cases result more accurate to obtain strong directional component, since the abrasion wear test establishes a
automatic prediction over a large dataset of training images but with specific sliding direction. Then, those aligned scratches observed in the
the limitation of loss interpretability in the obtained results. worn surfaces could not be differentiated as the number of orientations
is varied. It must be mentioned that SEM caption is present in all the
2.4. Statistical validation framework images, from which no effect could be expected.

Taking into account that classification task is dependent of number 3.2. Cross validation analysis of classifiers
of training samples, a statistical cross validation strategy was herein
used to analyze the performance of the proposed approach at different k-fold cross validation was performed over the set of classifiers and
number of samples. Such analysis was extended to observe the perfor- using different HOG configurations. An average of the 50-fold was es-
mance of the HOG descriptor using different configuration, for instance, timated as a function of the parameters configuration, as well as the
the patch size, the number of oriented bins and the number of patches maximum value of accuracy. The cross validation results of classifiers in
used to normalize the descriptor. Also, the cross validation was eval- terms of mean accuracy (%) for the different combinations of HOG
uated over different implemented classifiers. parameters are presented in Fig. 6.
The cross validation herein implemented was the k-fold strategy, in From a general point of view, all the classifiers had an outstanding
which the whole dataset is split iteratively in two complementary behavior since they predict very well the kind of wear exhibited by a
subsets. At each split one subset is called training set and the other is particular image. In other words, they can recognize an image ex-
named testing set to validate the particular approach. Such partition is hibiting mild and severe wear features in an accuracy of approximately
performed in multiple rounds to reduce variability in results and higher than 80%. The RaF classifier result in the most stable values of
properly characterize the predictive performance of the model. This accuracy. However, the DT classifier reached higher values close to
statistical evaluation, among others, allows to identify overfitting in 95% in some HOG configurations. It must be noted that the DT classifier
trained models and analyze the capability of generalization of classifi- displayed an unstable performance due to the high dispersion of ac-
cation strategies. Particularly, for our experiments the whole dataset is curacy values. The Gauss classifier had a stable behavior but in several
initially randomly shuffled and then partitioned into k equal sub- configurations showed low values of accuracy.
samples. Then k − 1 subsets are used to built a training model and the The maximum values of accuracy for the same configurations of
leftover subset is used for test the strategy. This process is repeated k HOG parameters are showed in Fig. 7. A very similar behavior was
times to ensure all possible samples into training and test subsets. noted between the maximum and mean values of accuracy. It must be
Finally, an average is computed over all k experiments to obtain the highlighted that most part of the RaF maximum values are grater than
accuracy of the particular approach. In a particular configuration of 90%, which gives a reliable result of the classifier capacity. At the same
k = n , with n the number of total images we obtain a leave-one-out cross- time, an accuracy of 100% can also be reached with the DT classifier,
validation in which the major number of “severe” and “mild” samples taking into account a determined 32 − c 64 − b 42 combination of HOG
are used to built the training model and only one image is tested at each parameters.
round. The performance of the HOG descriptor computation together To better analyze the effect of the number of grouped patches and
with classifier models was evaluated in terms of the classical statistical the patch size, results of accuracy are presented in Tables 3 and 4, re-
accuracy. Let TP , TN , FP and FN be the number of true positive, true spectively. Mean values, with the associated standard deviations in
negative, false positive and false negative, respectively, associated to parenthesis, and maximum values of accuracy are reported in both
each mild or severe labels. At each experiment, the annotated labels Tables. Initially, in Table 3 are shown the results in accuracy of clas-
could take values of True or False, while the predicted labels could be sifiers for three HOG configurations varying the number of grouped
positives or negatives. Hence, the accuracy is defined in Eq. (8): patches. There, the best performance of the DT classifier can be ob-
served, which obtained the highest accuracies in both mean and max-
TP + TN
Acc = imum values. In fact, for 32 orientations computed in patches with
TP + TN + FP + FN (8)
dimensions of 64 × 64 and using 42 patches for the block normalization
(32 − c 64 − b 42 HOG configuration), a maximum accuracy of 100%
3. Results and discussion was acquired.
In Table 4 the results obtained for each classifier are summarized,
3.1. HOG descriptor where the patch size was changed while keeping constant the other
HOG parameters. In general, for Gauss and DT classifiers the increase in
The results of the HOG computation with sample images of mild and the patch size from 16 to 32 results in an improvement of the de-
severe wear are showed in Fig. 5. This figure presents the images of scription of gradient patterns, achieving better classification accuracy.
worn surfaces with mild (a) and severe (b) wear, together with the The RaF performance is stable for all the patch size configurations.
corresponding HOG results with 4 − c16 − b1 (c) and 32 − c 64 − b1 (d) The effect of the number of k-folds on the accuracy values for the

1707
C. Gonzalez-Arias et al. Wear 426–427 (2019) 1702–1711

Fig. 5. HOG results for the images of worn surfaces with mild and severe wear and different configurations.

Fig. 6. Results of mean values in accuracy of classifiers for different HOG parameters configurations.

1708
C. Gonzalez-Arias et al. Wear 426–427 (2019) 1702–1711

Fig. 7. Results of maximum values in accuracy of classifiers for different HOG parameters configurations.

Table 3 were observed. For instance, the lower patch size results in lower values
Results of accuracy (%) for the classifiers with different HOG orientations. of accuracy and changed the best classifier model, from DT to RaF. In
Classifier % of accuracy with different HOG orientations
addition, the decrease in the patch size also moved the maximum values
of accuracy to higher numbers of k-fold for Gauss and RaF classifiers.
Config 8-c64-b4 16-c64-b4 32-c64-b4 From this viewpoint, smaller patch size contributes to acquire more
precise and higher values of accuracy and to have maximum values of
Mean Max. Mean Max. Mean Max.
accuracy at low numbers of k-fold. In other words, an image analysis of
Gauss 0.88 , (0.22 ) 0.90 0.87 , (0.23 ) 0.90 0.86 , (0.23 ) 0.91 smaller regions makes possible to use few images to build the training
RaF 0.85 , (0.23 ) 0.92 0.85 (0.24 ) 0.92 0.86 , (0.23 ) 0.94 model of each classifier and thus the computational cost of analysis is
DT 0.91, (0.18 ) 0.97 0.94 (0.15) 0.98 0.96 , (0.13 ) 1 reduced.
Some limitations are reported for the proposed work, when image
samples of different classes report gradient patterns very similar (see
Table 4
Fig. 2, examples (b) and (f)). In such cases, the configuration of HOG
Results of accuracy (%) for the classifiers with different HOG patch sizes.
descriptor with few orientations, i.e., {8, 16} , produce a miss-classifi-
Classifier % of accuracy with different HOG patch sizes cation. This fact is associated to the poor representation of orientations
that only recovers strong directions of gradients. Interestingly enough,
Config 16-c16-b4 16-c32-b4 16-c64-b4
for HOG configuration with histograms of 64 bin orientations and patch
Mean Max. Mean Max. Mean Max. size of c = 64 , the proposed strategy achieves a proper classification for
these kind of examples, using the DT and RaF algorithms. The addi-
Gauss 0.67 , (0.32 ) 0.71 0.87 , (0.23 ) 0.91 0.87 , (0.23 ) 0.90 tional direction on histograms allows to enrich the representation of
RaF 0.85 , (0.24 ) 0.95 0.87 , (0.23 ) 0.96 0.85 , (0.24 ) 0.92
surfaces and differentiate very local patterns in regions. Also the block
DT 0.72 , (0.29 ) 0.82 0.83 , (0.25 ) 0.90 0.94 , (0.15 ) 0.98
normalization contributes to a smoothing of small bins that much of the
time represent noise on images. In such configuration, for some ex-
three classifier models in determined HOG configurations is displayed periments, the proposed approach achieves a perfect classification,
in Fig. 8, where in addition the error bars and the maximum values of demonstrating the robustness of descriptor to represent wear regimes.
accuracy (vertical) for each classifier are shown. In the 16-c64-b42
configuration it can be observed that all the classifier models had a
4. Conclusions
stable behavior, as can be seen in Fig. 8(a), and the standard deviation
increased with the number of k-folds. Generally, highest values of ac-
In this work a geometrical image descriptor (HOG) together with a
curacy were obtained for the DT classifier. It must be noted that max-
classifier model were used to accurately classify SEM images of worn
imum values of accuracy were obtained at low values of k-fold. The
surfaces exhibiting mild and severe abrasive wear regimes. Three
results show that a cross validation with low k-folds, for a certain HOG
classifier models were statistically tested by means of a cross validation
configuration, which could lead to acquire minimum dispersion of data
process to evaluate their performance. The main contribution of this
and maximum values of accuracy for the three classifier models. This
work was the automatic abrasive wear regime recognition from gra-
would be advantageous in terms of computational cost.
dient characteristics in worn surfaces images.
On the other hand, the results in Fig. 8 allow knowing the effect of
The novel computational strategy was able to index wear severity
decreasing the patch size while keeping the other HOG parameters
from worn surfaces images. The performance of the RaF classifier was
constant. Initially, it can be observed how the best results are reached
outstanding to describe the type of wear obtained in SEM images of
with the RaF classifier but with lower accuracies than those obtained by
worn surfaces, represented with mean values of prediction accuracy
the DT classifier in the previous HOG configuration. Again, at low va-
higher than 80% in most part of the HOG computation. The RaF dis-
lues of k-folds the standard deviation for all models was reduced.
played the most consistent behavior without considering the HOG
However, in this HOG configuration maximum values of accuracy for
configuration, but in some particular HOG configurations was sur-
Gauss and RaF models were acquired at high values of k-folds, i.e., in
passed by the DT classifier. The DT classifier had an outstanding be-
the particular case when much of the images are used to built the
havior in which eventually could predict the type of wear regime from
training model of each classifier
an image with a high accuracy. However, generally this classifier ex-
Some dependencies of classifiers accuracy on the HOG parameters
hibited an irregular performance with some low mean and maximum

1709
C. Gonzalez-Arias et al. Wear 426–427 (2019) 1702–1711

Fig. 8. Results of accuracy in terms of the number of k-folds for the 16-c64-b42 (a) and the 16-c16-b42 configurations.

values of accuracy for some HOG configurations. In conclusion, an References


optimization must be carried out to find specific HOG configurations in
which a high accuracy with a low computational cost is obtained for the [1] I. Hutchings, P. Shipway, Tribology - Friction and Wear of Engineering Materials,
DT classifier. 2nd ed., Elsevier, Butterworth-Heinemann, 2017.
[2] K. Holmberg, Reliability aspects of tribology, Tribol. Int. 34 (12) (2001) 801–808,
From the viewpoint of future work, the proposed strategy con- https://doi.org/10.1016/S0301-679X(01)00078-0.
stitutes a basis for an online monitoring and detection system to eval- [3] H. Spikes, Tribology research in the twenty-first century, Tribol. Int. 34 (2001)
uate wear severity of worn surfaces. However, it is necessary to carry 789–799, https://doi.org/10.1016/S0301-679X(01)00079-2.
[4] J. Zhang, Detection and Monitoring of Wear Using Imagingmethods (Ph.D. thesis),
out an analysis of images of worn surfaces of severe and mild abrasive University of Twente, 2006.
wear regimes obtained at low magnifications, for example, by means of [5] F. Giusti, M. Santochi, G. Tantussi, On-line sensing of flank and crater wear of
a video camera. This would be fundamental to develop a detection cutting tools, CIRP Ann. - Manuf. Technol. 36 (1) (1987) 41–44, https://doi.org/10.
1016/S0007-8506(07)62549-9.
system to have an online automated recognition of abrasive wear re- [6] J.U. Jeon, S.W. Kim, Optical flank wear monitoring by image processing, Wear 127
gimes in machine parts surfaces. This analysis could be applied also to (1988) 207–217, https://doi.org/10.1016/0043-1648(88)90131-7.
the sliding wear of metals, where worn surfaces with severe and mild [7] K.B. Pedersen, Wear measurement of cutting tools by computer vision, Int. J. Mach.
Tools Manuf. 30 (1) (1990) 131–139, https://doi.org/10.1016/0890-6955(90)
wear regimes also display well-defined features [23]. In fact, the
90047-M.
computational strategy here proposed could also be applied in other [8] S. Kurada, C. Bradley, A machine vision system for tool wear assessment, Tribol. Int.
types of wear, as long as transitions between different wear regimes can 30 (4) (1997) 295–304, https://doi.org/10.1016/S0301-679X(96)00058-8.
be observed. [9] M. Yang, O. Kwon, Crater wear measurement using computer vision and automatic
focusing, J. Mater. Process. Technol. 58 (1996) 362–367, https://doi.org/10.1016/
0924-0136(95)02208-2.
[10] Y. Maeda, H. Uchida, A. Yamamoto, Measurement of the geometric features of a
Acknowledgements cutting tool edge with the aid of a digital image processing technique, Precis. Eng.
11 (3) (1989) 165–171, https://doi.org/10.1016/0141-6359(89)90071-8.
[11] J. Zhou, Q. Zhou, L. Yang, G. Wu, Acquisition and rendering of characteristic
The authors greatly acknowledge the financial support offered by images of the wear surfaces of running parts within machinery, Wear 265 (3–4)
the Universidad Industrial de Santander (UIS) through the project VIE- (2008) 540–545, https://doi.org/10.1016/j.wear.2007.09.002.
[12] S. Soleimani, J. Sukumaran, A. Kumcu, P. De Baets, W. Philips, Quantifying abra-
2303. sion and micro-pits in polymer wear using image processing techniques, Wear 319
(1–2) (2014) 123–137, https://doi.org/10.1016/j.wear.2014.07.018.

1710
C. Gonzalez-Arias et al. Wear 426–427 (2019) 1702–1711

[13] A.D.H. Thomas, T. Davies, A.R. Luxmoore, Computer image analysis for identifi- 212–232, https://doi.org/10.1016/j.cirpj.2013.02.005.
cation of wear particles, Wear 42 (1991) 213–226, https://doi.org/10.1016/0043- [23] C.C. Viáfara, A. Sinatora, Influence of hardness of the harder body on wear regime
1648(91)90165-Q. transition in a sliding pair of steels, Wear 267 (1–4) (2009) 425–432, https://doi.
[14] Z. Jin, H. Shao, Computer image analysis of abrasion debris, Wear 154 (1) (1992) org/10.1016/j.wear.2008.11.019.
1–10, https://doi.org/10.1016/0043-1648(92)90238-4. [24] G. Pintaude, F.G. Bernardes, M.M. Santos, A. Sinatora, E. Albertin, Mild and severe
[15] I. Bodini, C. Petrogalli, M. Faccoli, M. Lancini, S. Pasinetti, G. Sansoni, F. Docchio, wear of steels and cast irons in sliding abrasion, Wear 267 (1–4) (2009) 19–25,
A. Mazzù, Evaluation of wear in rolling contact tests by means of 2D image analysis, https://doi.org/10.1016/j.wear.2008.12.099.
Wear 400–401 (2018) 156–168, https://doi.org/10.1016/j.wear.2017.12.023. [25] J.J. Coronado, S.A. Rodríguez, A. Sinatora, Effect of particle hardness on mild-se-
[16] T.B. Kirk, D. Panzera, R.V. Anamalay, Z.L. Xu, Computer image analysis of wear vere wear transition of hard second phase materials, Wear 301 (1–2) (2013) 82–88,
debris for machine condition monitoring and fault diagnosis, Wear 181–183 (1995) https://doi.org/10.1016/j.wear.2012.12.016.
717–722, https://doi.org/10.1016/0043-1648(95)90188-4. [26] K. Holmberg, A. Helle, Tribology as basis for machinery condition diagnostics and
[17] J. Rawers, J. Tylczak, Fractal characterization of wear-erosion surfaces, J. Mater. prognostics, Int. J. Perform. Eng. 2000 (2) (2008) 255–269.
Eng. Perform. 8 (6) (1999) 669–676, https://doi.org/10.1361/ [27] R. Lewis, U. Olofsson, Mapping rail wear regimes and transitions, Wear 257 (7–8)
105994999770346431. (2004) 721–729, https://doi.org/10.1016/j.wear.2004.03.019 (arXiv:1011.
[18] A. Abouel-Kasem, M.A. Al-Bukhaiti, K.M. Emara, S.M. Ahmed, Fractal character- 1669v3 ).
ization of slurry eroded surfaces at different impact angles, J. Tribol. 131 (3) (2009) [28] P. Kittipanya-ngam, E.H. Lung, HOG-based descriptors onrotation invariant human
31601, https://doi.org/10.1115/1.3118784. detection, in: R. Koch, F. Huang (Eds.), Computer Vision – ACCV 2010 Workshops,
[19] Y.C. Tasan, M.B. De Rooij, D.J. Schipper, Measurement of wear on asperity level Lecture Notes in Computer Science, Springer, Berlin, Heidelberg, 2010, pp.
using image-processing techniques, Wear 258 (1–4) (2005) 83–91, https://doi.org/ 143–152, , https://doi.org/10.1007/978-3-642.
10.1016/j.wear.2004.05.018. [29] P.J. Bickel, E. Levina, Some theory for Fisher's linear discriminant function, naive
[20] B. Meylan, P. Dogan, D. Sage, K. Wasmer, A simple, fast and low-cost method for in Bayes, and some alternatives when there are many more variables than observa-
situ monitoring of topographical changes and wear rate of a complex tribo-system tions, Bernoulli 10 (6) (2004) 989–1010, https://doi.org/10.3150/bj/1106314847.
under mixed lubrication, Wear 364–365 (2016) 22–30, https://doi.org/10.1016/j. [30] R. Tibshirani, T. Hastie, Margin trees for high-dimensional classification, J. Mach.
wear.2016.06.006. Learn. Res. 8 (2007) 637–652.
[21] J. Sukumaran, S. Soleimani, P. De Baets, V. Rodriguez, K. Douterloigne, W. Philips, [31] W. Loh, N. Vanichsetakul, Tree-structured classification via generalized dis-
M. Ando, High-speed imaging for online micrographs of polymer composites in criminant analysis, J. Am. Stat. Assoc. 83 (1988) 715–725.
tribological investigation, Wear 296 (1–2) (2012) 702–712, https://doi.org/10. [32] T.K. Ho, The random subspace method for constructing decision forests, IEEE Trans.
1016/j.wear.2012.08.016. Pattern Anal. Mach. Intell. 20 (8) (1998) 832–844, https://doi.org/10.1109/34.
[22] S. Dutta, S.K. Pal, S. Mukhopadhyay, R. Sen, Application of digital image processing 709601.
in tool condition monitoring: a review, CIRP J. Manuf. Sci. Technol. 6 (3) (2013)

1711

You might also like